Edge Classification Based Initialization Technique for Image Vector Quantizer Design.

Xiaodan Jiang,Zhe-Ming Lu

2018-01-01

Abstract:. Codebook design is one of key techniques in Vector Quantization (VQ). The K-means algorithm is the most widely used algorithm to design a VQ codebook. However, both convergence speed and performance of the generated codebook depend on the initial codebook. Thus, many algorithms have been presented to obtain a good initial codebook, including the well known splitting, pruning, pairwise nearest neighbor design, random initialization and maximum distance initialization. However, all these algorithms do not make full use of the features of the training vectors to obtain a better initial codebook and some of them need high extra overhead. This paper provides a very simple technique to obtain a better initial codebook by classifying training vectors according to their edge orientations dependent on eight edge templates proposed by us, and thus we can obtain eight subsets. Then we randomly select codewords from each subset with the number of codewords being proportional to the number of training vectors in the subset. Experimental results show that, compared with the conventional and modified K-means algorithms based on random selection initialization, our initialization technique converges to a better locally optimal codebook with a faster convergence speed.